Channel Spatial Attention-based Transformer for Image Super-Resolution

NCU Institutional Repository > 資訊電機學院 > 資訊工程研究所 > 博碩士論文 > Item 987654321/93537

jsp.display-item.identifier=請使用永久網址來引用或連結此文件: http://ir.lib.ncu.edu.tw/handle/987654321/93537

题名:	Channel Spatial Attention-based Transformer for Image Super-Resolution
作者:	陳家偉;CHEN, CHIA-WEI
贡献者:	資訊工程學系
关键词:	超解析度;Super Resolution;Transformer
日期:	2024-01-25
上传时间:	2024-09-19 17:11:35 (UTC+8)
出版者:	國立中央大學
摘要:	隨著影音媒體需求的不斷增長，超解析度領域的重要性日益提升。特別是Transformer模型因其卓越的性能而在電腦視覺方面受到廣泛關注，導致越來越多的研究將其應用於這一領域。然而，我們發現儘管Transformer通過增加不同機制的注意力能夠解決學習特徵有限的問題，但是在訓練過程中仍可能遺失一些紋理和結構。為了盡可能地保留初始特徵和架構，我們提出了一種方式用以整合Residual Connection、Attention Mechanism和Upscaling Technique。為了驗證我們方法的性能，我們在五個不同的資料集上進行了多次實驗，並且與現有的先進超解析度模型進行了比較。實驗結果顯示，我們的方法在性能上相較於當前領域中最先進的模型有著更佳的表現。;As the demand for audio-visual media continues to grow, the significance of the super-resolution field is increasingly recognized. In particular, Transformer models have garnered widespread attention in the realm of computer vision due to their exceptional performance, leading to their growing application in this area. However, we observed that despite the ability of Transformer to address the issue of limited feature learning through various attention mechanisms, some textures and structures may be lost during the learning process. To maximally preserve the initial features and structures, we propose a system, named Integrated Attention Transformer (IAT), that integrates Residual Connection, Attention Mechanism, and Upscaling Technique. To confirm the efficacy of IAT, experiments were conducted on five different datasets, compared with the current advanced super-resolution state-of-the-art (SOTA) model. The results show that the proposed IAT surpasses the current SOTA model.
显示于类别:	[資訊工程研究所] 博碩士論文

文件中的档案:

档案	描述	大小	格式	浏览次数
index.html		0Kb	HTML	10	检视/开启

在NCUIR中所有的数据项都受到原著作权保护.

社群 sharing

数据加载中.....